Ranked Relevance Results Using Multi-Feature Scoring Returned from a Universal Relevance Service Framework

1. A computer-implemented method for calculating a user-deal relevance score, the method comprising: generating one or more trained machine learning models based at least in part on supervised learning models and training data, the one or more trained machine learning models configured for generating a user-deal relevance score based at least in part on one or more feature vectors associated with user-deal interaction data associated with a particular user; receiving, using at least one processor and via a relevance service application programming interface (API) from a relevance API client device, a search request generated on behalf of the particular user for deals describing promotion offerings that are currently available to the particular user; in response to receiving the search request, retrieving, using the at least one processor and from a data store, a set of join tables including user-interaction data describing user-deal interactions for the particular user and a particular deal; transforming, using the at least one processor, the set of join tables into a plurality of feature vectors; calculating, using the at least one processor and the set of join tables, a set of features including odds, DDO, and similarity, wherein odds comprises historical conversion data computed from jointed user and deal attributes, wherein DDO comprises past deal performance based on mapping an individual deal's performance data to all users similar to the particular user, and wherein similarity comprises a likelihood that the particular user is going to purchase deals similar to the particular deal; generating, using the at least one processor, the user-deal relevance score based at least in part on applying the one or more trained machine learning models to the plurality of feature vectors and on the set of features, wherein calculating the user-deal relevance score is based on a linear model derived from similarity, blended odds, and DDO features, and wherein calculating the blended odds and DDO features is based on Bayesian blending of odds and DDO features in user-deal space; and in an instance where a ratio of a click likelihood for the particular user to an average click likelihood exceeds a predefined number, adjusting, using the at least one processor, the user-deal relevance score based at least in part on the ratio; generating, using the at least one processor and based at least on the adjusted user-deal relevance score, a list of ranked deals for presentation via an electronic interface of a user device associated with the particular user; and transmitting, using the at least one processor and to the user device, one or more deals from the list of ranked deals configured for display within the electronic interface of the user device.

2. The method of claim 1 , wherein the calculating of the blended odds and DDO features based on Bayesian blending of odds and DDO in user-deal space comprises: estimating, using the at least one processor, n prior events were generated by users during a particular time period; collecting, using the at least one processor, m actual events that were generated by the particular user during the particular time period; and calculating, using the at least one processor, a posterior estimation based on a weighted average of the n prior events and the m actual events.

3. The method of claim 2 , further comprising: in an instance in which there is insufficient data to enable odds calculation, using the at least one processor, rolling up the data to a higher level of organization.

4. The method of claim 3 , wherein rolling up the data to a higher level of organization follows the structure of an odds decision tree.

5. The method of claim 2 , further comprising: for each of the particular user's attributes, determining, using the at least one processor, whether there is sufficient data describing the attribute to calculate DDO; and using the attribute to calculate, using the at least one processor, DDO in an instance in which there is sufficient data describing the attribute.

6. The method of claim 5 , wherein determining whether there is sufficient data describing the attribute to calculate DDO comprises: using, by the at least one processor, a Two-One-Sided-Test (TOST) to test equivalence.

7. The method of claim 1 , wherein data describing user and deal attributes are represented in a set of join tables, and wherein calculating similarity for the particular user and the particular deal comprises: generating, using the at least one processor, an extended user attribute vector in user-deal space, wherein the generating is based on a user vector representing the particular user's attributes and a projected user vector representing the user vector projection into deal space; generating, using the at least one processor, an extended deal attribute vector in user-deal space, wherein the generating is based on a deal vector representing the particular deal's attributes and a projected deal vector representing the deal vector projection into user space; and calculating, using the at least one processor, similarity for the particular user and the particular deal by taking the dot product of the extended user attribute vector and the extended deal attribute vector.

8. The method of claim 7 , wherein calculating the user-deal relevance score is based on incorporating DDO and odds into similarity, and wherein incorporating DDO and odds into similarity comprises: generating, using the at least one processor, user-deal space evidence based on a projection of the particular user's attributes into deal space; blending, using the at least one processor, the user-deal space evidence with the projected user vector; generating, using the at least one processor, deal-user space evidence based on a projection of the particular deal's attributes into user space; blending, using the at least one processor, the deal-user space evidence with the projected deal vector; and calculating, using the at least one processor, similarity for the particular user and the particular deal by taking the dot product of the blended vectors.

9. The method of claim 1 , wherein calculating the user-deal relevance score is based on multiple models including at least one of a linear model and a machine-learning model.

10. A computer program product, stored on a computer readable medium, comprising instructions that when executed on one or more computers cause the one or more computers to: generate one or more trained machine learning models based at least in part on supervised learning models and training data, the one or more trained machine learning models configured for generating a user-deal relevance score based at least in part on one or more feature vectors associated with user-deal interaction data associated with a particular user; receive, via a relevance service application programming interface (API) from a relevance API client device, a search request generated on behalf of a particular user for deals describing promotion offerings that are currently available to the particular user; in response to receiving the search request, retrieve, from a data store, a set of join tables including user-interaction data describing user-deal interactions for the particular user and a particular deal; transform, the set of join tables into a plurality of feature vectors; calculate, using the set of join tables, a set of features including odds, DDO, and similarity, wherein odds comprises historical conversion data computed from jointed user and deal attributes, wherein DDO comprises past deal performance based on mapping an individual deal's performance data to all users similar to the particular user, and wherein similarity comprises a likelihood that the particular user is going to purchase deals similar to the particular deal; generate the user-deal relevance score based at least in part on applying the one or more trained machine learning models to the plurality of feature vectors and on the set of features, wherein calculating the user-deal relevance score is based on a linear model derived from similarity, blended odds, and DDO features, and wherein calculating the blended odds and DDO features is based on Bayesian blending of odds and DDO features in user-deal space; in an instance where a ratio of a click likelihood for the particular user to an average click likelihood exceeds a predefined number, adjust the user-deal relevance score based at least in part on the ratio; generate, based at least on the adjusted user-deal relevance score, a list of ranked deals for presentation via an electronic interface of a user device associated with the particular user; and transmit, to the user device, one or more deals from the list of ranked deals configured for display within the electronic interface of the user device.

11. The computer program product of claim 10 , wherein calculating the user-deal relevance score is based on a linear model of similarity, odds, and DDO features.

12. The computer program product of claim 10 , wherein the calculating of the blended odds and DDO features based on Bayesian blending of odds and DDO in user-deal space comprises: estimating n prior events were generated by users during a particular time period; collecting m actual events that were generated by the particular user during the particular time period; and calculating a posterior estimation based on a weighted average of the n prior events and the m actual events.

13. The computer program product of claim 12 , wherein the one or more computers is further caused to: in an instance in which there is insufficient data to enable odds calculation, roll up the data to a higher level of organization.

14. The computer program product of claim 13 , wherein rolling up the data to a higher level of organization follows the structure of an Odds decision tree.

15. The computer program product of claim 12 , wherein the one or more computers is further caused to: for each of the particular user's attributes, determine whether there is sufficient data describing the attribute to calculate DDO; and use the attribute to calculate DDO in an instance in which there is sufficient data describing the attribute.

16. The computer program product of claim 15 , wherein determining whether there is sufficient data describing the attribute to calculate DDO comprises: using a Two-One-Sided-Test (TOST) to test equivalence.

17. The computer program product of claim 10 , wherein data describing user and deal attributes are represented in a set of join tables, and wherein calculating similarity for the particular user and the particular deal comprises: generating an extended user attribute vector in user-deal space, wherein the generating is based on a user vector representing the particular user's attributes and a projected user vector representing the user vector projection into deal space; generating an extended deal attribute vector in user-deal space, wherein the generating is based on a deal vector representing the particular deal's attributes and a projected deal vector representing the deal vector projection into user space; and calculating similarity for the particular user and the particular deal by taking the dot product of the extended user attribute vector and the extended deal attribute vector.

18. The computer program product of claim 17 , wherein calculating the user-deal relevance score is based on incorporating DDO and odds into similarity, and wherein incorporating DDO and odds into similarity comprises: generating user-deal space evidence based on a projection of the particular user's attributes into deal space; blending the user-deal space evidence with the projected user vector; generating deal-user space evidence based on a projection of the particular deal's attributes into user space; blending the deal-user space evidence with the projected deal vector; and calculating similarity for the particular user and the particular deal by taking the dot product of the blended vectors.

19. The computer program product of claim 10 , wherein calculating the user-deal relevance score is based on multiple models including at least one of a linear model and a machine-learning model.